Scope 3, meet Claude
Is AI in your sustainability report?
Take a look at my website!
Dear Sustainability Manager,
Let’s be honest here, your colleagues are using AI. Your procurement team is running supplier queries through ChatGPT, your marketing department is generating copy with Copilot and someone in finance is definitely asking Claude to clean up their Excel formulas. AI has quietly become part of how work gets done and it is time to think what it means for your sustainability report.
According to multiple workforce surveys from 2024-2025, the majority of knowledge workers use AI tools regularly often regardless of whether their employer has an official policy. We’re not talking about a niche tech team habit. We’re talking about AI-assisted drafting, data analysis, image generation, code writing and customer service happening at scale, across industries, every single day.
For a sustainability manager, this is interesting for two reasons:
Energy consumption: AI inference (actually running a query) and training both consume significant electricity. A single query to a large language model uses roughly 10× the energy of a standard Google search and that gap widens for image generation or complex reasoning tasks.
Data center footprint: the servers running these models are thirsty for both power and water. Microsoft, Google and Amazon have all flagged increasing data center energy demand as a material challenge to their own net-zero commitments.
So when your team uses AI tools, there is a real emissions footprint attached. The question is: whose emissions are they and how do you account for them?
Scope 1, 2 or 3? Where does AI energy use actually sit?
This is where it gets genuinely tricky and where a lot of sustainability managers are currently underreporting without realising it.
If you run AI on your own infrastructure (on-premise or private cloud)
The electricity powering your servers falls under Scope 2, purchased electricity consumed by your operations. This is relatively straightforward: you measure the electricity used by your data centers or server rooms and apply the appropriate emission factor for your grid. If you’re already doing good Scope 2 accounting, AI workloads are just another load on that infrastructure.
If you use cloud-based AI services (which is most of us)
Here the picture shifts. When you pay for API calls to OpenAI, Microsoft Azure AI or Google Vertex, you’re consuming a service and the emissions from producing that service land in Scope 3, Category 1 (Purchased Goods and Services).
Under the GHG Protocol, services you procure from third parties are Scope 3. Full stop. The electricity that OpenAI or Anthropic uses to run your queries is not your Scope 2, it’s their Scope 1/2 and your Scope 3.
Practical implication: if your company has set a Scope 3 target (which is increasingly expected under SBTi and CSRD frameworks), AI service usage should be in scope literally and figuratively.
What about Microsoft Copilot bundled in your M365 subscription?
This is a real grey zone that’s catching organisations off guard. If AI functionality is embedded in software you already license (Teams, Office, Salesforce etc.), the incremental emissions are easy to overlook. But they exist and as these features get used more heavily, the footprint grows. The principle is the same: it’s Scope 3, Category 1.
How to account for it in Sustainability reports and LCAs
Here’s the honest answer: the methodologies are still maturing, but that doesn’t mean you do nothing. Below you will find a pragmatic approach.
1. Establish a baseline
Start with what you can measure or estimate:
Volume of AI API calls per month (check your billing dashboards, OpenAI, Azure, Google Cloud all report token usage)
Type of tasks: text inference is much lighter than image generation or model fine-tuning
Which tools your teams are using (a quick internal survey is your friend here)
2. Apply emission factors
This is the hard part, because AI providers are inconsistent in disclosing energy-per-inference data. Your options:
Use provider sustainability reports: Microsoft, Google, and Amazon publish annual sustainability reports with data center PUE (Power Usage Effectiveness) and energy mix data. Some are beginning to disclose AI-specific metrics.
Use proxy figures: researchers like those at the ML CO2 Impact project and Emma Strubell’s landmark work on NLP energy use provide order-of-magnitude estimates. For a rough LCA, using ~0.002-0.01 kWh per text query (depending on model size and provider efficiency) is a defensible starting point with appropriate uncertainty documentation.
Ask your vendors: under CSRD and supplier engagement expectations, you have every right to request emissions data from your AI service providers. Some, like Google Cloud, already offer carbon footprint tools in their console.
3. Document in your sustainability report
When disclosing:
Classify correctly: Scope 3, Category 1 for cloud AI services
Be transparent about methodology: State your assumptions, data sources and uncertainty ranges, auditors and ratings agencies appreciate honesty over false precision
Flag materiality: is AI usage currently immaterial relative to your overall footprint? Say so, and commit to monitoring as usage grows
Trend over time: even if the absolute number is small today, showing year-on-year tracking demonstrates governance maturity
4. For LCA specifically
If you’re conducting a Life Cycle Assessment for a product or service that uses AI in its production or delivery (increasingly common in software products, data analytics services or AI-assisted manufacturing), you’ll want to:
Include AI compute as an input in the use phase or production phase inventory
Use ecoinvent or similar LCA databases for electricity inputs, applying your vendor’s disclosed energy mix where possible
Consider the functional unit carefully: is the AI usage proportional to output units or is it a fixed overhead?
The ISO 14040/14044 framework doesn’t specifically address AI yet, but the underlying principle applies cleanly.
Sustainability managers are used to navigating imperfect data and evolving standards. AI emissions accounting is no different from where Scope 3 Category 11 (use of sold products) was a decade ago: messy, under-standardised, but increasingly important.
The organisations that start measuring now, even roughly, will be far better positioned when regulators and frameworks catch up and given the trajectory of AI adoption, they will catch up.
In the meantime: ask your IT team for AI tool usage data, start a conversation with your cloud vendors, and make sure your next sustainability report at least acknowledges the question. That’s already ahead of most.
See you next week!
Gianluca


Loved it, super interesting!
the right tool to address AI emissions is SCI for AI from the GreenSoftware.foundation, it splits AI usage in provider and consumer and it gives a methodology to calculate the impacts/emissions.